基于语义分割的室外场景识别技术研究

Research on Outdoor Scene Recognition Technology Based on Semantic Segmentation

针对于场景识别问题,提出一种基于开源的室外场景数据集以及自定义采集的数据集在deepLabV3+深度学习模型上进行实验,并运用一种改进的K-近邻算法对DeepLabV3+深度学习模型进行优化。与现有的测试数据集的方法不同,省去了对数据集进行标签的工作,减少了大量的前期准备工作,提高了模型的计算效率以及分类模型的准确率和召回率。结果表明,使用K-近邻算法改进后的Deeplabv3+模型识别精度达到相0.75,较于直接使用Deeplabv3+模型进行语义分割的准确率0.65提高了0.1,并且得到了效果明显,在一定程度上提升了实验效率以及算法的鲁棒性。

Aiming at the problem of scene recognition, an open-source outdoor scene data set and a custom-collected data set were proposed to experiment on the DeepLabV3+deep learning model, and an improved K-nearest neighbor algorithm was used to optimize the DeepLabV3+deep learning model. Different from the existing method of testing the data set, the work of labeling the data set was omitted, a lot of preliminary preparation work was reduced, and the calculation efficiency of the model and the accuracy and recall rate of the classification model were improved. The results show that the recognition accuracy of DeepLabV3+model improved by k-nearest neighbor algorithm reaches 0.75,which is 0.1 higher than that of 0.65 for semantic segmentation directly using DeepLabV3+model, and the effect is obvious, which improves the experimental efficiency and the robustness of the algorithm to a certain extent.